Adaptive inference

Adaptive inference is the process by which an observer — whether biological, computational, or social — updates not only its beliefs about the world but also the inferential framework through which those beliefs are formed. Unlike standard Bayesian inference, which assumes a fixed hypothesis space and a stationary prior, adaptive inference treats the hypothesis space itself as subject to revision when the evidence systematically fails to fit the available categories. The observer does not merely learn which hypothesis is true; it learns which hypotheses are even possible.

This process is central to observer selection: the criteria by which an observer distinguishes signal from noise are themselves inferred from the structure of past observations, and they evolve as the observer encounters regimes where the old criteria fail. Hierarchical Bayesian models provide one formalization of adaptive inference, in which higher-level priors encode beliefs about the structure of the lower-level hypothesis space. When the evidence is sufficiently anomalous, the higher-level prior shifts, producing a qualitative change in what the observer can perceive. This is not a bug in the inference process. It is the mechanism by which observers escape the informational boundaries of their own prior commitments.